2
$\begingroup$

I have a cross-sectional data set with about 8000 observations on child obesity (eg BMI). This data was collected in 8 countries and within schools (about 200 schools), i.e. observations are clustered within schools and countries. I am interested in how child characteristics (e.g. socio economic status) relate to child obesity. My approach was to estimate a pooled OLS and clustering at the school level and also including country dummy variables in the regressions. As I have many and reasonably large school clusters, I thought this would be a good approach. I have however been told that I should estimate a FE effect model. Is it possible to estimate a country fixed-effects model and cluster at the school level?

$\endgroup$
1
  • $\begingroup$ Since you have doubly nested data (students in schools in countries), I think you have to account for that in your model and use a 3 level multilevel model. $\endgroup$
    – Peter Flom
    Feb 13 '13 at 12:16
2
$\begingroup$

I think it is doable; this would be one of the relatively rare examples where the assumptions on the sample sizes will be met sufficiently well. Most education researchers would want to fit a classic multilevel model with three levels, and then I would say that 8 countries are not enough to reliably estimate variances at level 3. Depending on the software, you might be able to mix fixed effects with clustered standard errors (Stata xtreg ..., fe vce(cluster ...)), or you may have to include the country dummies as regressors to implement the country fixed effects.

You may also want to try some country-school or country-individual interactions, as TV programs in North Korea may not be so interesting to watch as in the US, so kids would spend less time in front of TV in some countries.

$\endgroup$
3
  • 1
    $\begingroup$ About your comment on TV programs in North Korea and the US... oh, never mind... $\endgroup$
    – FairMiles
    Feb 13 '13 at 19:42
  • $\begingroup$ Thank you very much for your kind responses. We have now tried a 3-level MLM (even with only 8 countries) and the results are very similar to OLS with multilevel clustering (or for that matter only clustering at the school level). We have also tried a country fixed effects model with clustering at the school level (xtreg ..., fe vce(cluster ...)), but that did not work. ...still trying to work that out. Anyone following this discussion and having the same problems may find this link useful: stata.com/meeting/13uk/nichols_crse.pdf $\endgroup$
    – Alfonso
    Feb 14 '13 at 10:37
  • $\begingroup$ Did not work as in, Stata reported an error, or as in, you did not like the results, or as in, there were no red errors but the standard errors were missing? Austin Nichols is worth listening to, although his talks are just too intense... too many words per slide. $\endgroup$
    – StasK
    Feb 14 '13 at 21:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.